Written by
递归客.
This journey was informed by the work of 递归客 (diguike), the author of the source book. Their analysis makes the design of agent memory something you can study, inspect, and build.
Agent Memory 工程实战：从 claude-mem 源码到企业级记忆平台
Agent Memory Engineering in PracticeEnglish rendering of the title for this companion.
What comes from whom
The source book and its engineering analysis: 递归客 (diguike). The original book is in Chinese and studies Claude-Mem v12.6.2. Its architectural explanations and code examples informed the topics taught here.
The guiding thought metaphor: @thedotmack, creator of Claude-Mem. His description of a few thoughts at a time in roughly three-second moments supplies the desk-and-notebook framing. This is a personal analogy, not a cognitive-science finding.
This interactive companion: original English explanations, fictional examples, calculations, illustrations, and browser experiments prepared for Claude-Mem. It does not reproduce or translate the book’s prose or code. Attribution to the source does not imply the author’s endorsement of this companion.
Respecting the author’s license
The source book’s prose is licensed under Creative Commons Attribution–NonCommercial–ShareAlike 4.0. Its code examples are separately licensed under MIT, as stated in the author’s README.
The book asks readers to credit the author and link to inferloop.dev when sharing it, and to contact the author for commercial use. That license remains with the original. A link or attribution does not grant permission to commercially reprint or translate it.
This course teaches the underlying ideas using its own wording and examples. Read the original license and copyright statement before reusing material from the book itself.
What the experiments can tell you
Every experiment is an illustrative simulation with invented event-planning data. None reads a personal notebook, captures your activity, calls a model, or implements a production memory service. Search uses a tiny hand-authored example; token amounts are chosen for teaching rather than measured with a tokenizer.
The thought calculator shows capacity arithmetic. It does not measure reasoning quality, compress your own data, or promise a particular number of memories for a model. Context-window size, useful note size, retrieval quality, and task performance need separate evaluation.
Completion marks remain only during your current visit. No account is needed for the course.
Take the next step with the original.
- 01The memory problem
- 02Claude-Mem at a glance
- 03Managing the context window
- 04System architecture
- 05Events and the session lifecycle
- 06The background worker
- 07Storage and search
- 08Progressive disclosure
- 09MCP search
- 10Observations
- 11The knowledge agent
- 12Development setup
- 13Building mini-mem
- 14Extending the system
- 15Other approaches to memory
- 16From plugin to platform
- 17Team memory and governance
- 18What comes next
Chapter labels are brief English descriptions. The author’s original chapter titles and text are available through the links.
Source revision:27a3d8f4c360. The book presents later platform and frontier chapters as design exploration; this course preserves that distinction.